Reproduce “Diffusion-based Inverse Model of a Distributed Tactile Sensor for Object Pose Estimation” on OpenResearch

About this paper

Ante Marić, Giammarco Caroleo, Alessandro Albini, Julius Jankowski, Perla Maiolino, Sylvain Calinon

A framework for vision-free object pose estimation leverages a diffusion-based inverse model to generate multimodal pose hypotheses from distributed tactile sensor data. This approach, combined with geometry-aware inference and a particle filter, achieves robust online pose estimation and tracking, demonstrating reliable convergence and accuracy in both simulated and real-world environments without strong initial priors.

Key insights

• Denoising diffusion models can effectively learn to generate diverse, high-likelihood pose hypotheses from ambiguous tactile data, serving as an efficient inverse sensor model.
• Combining a learned generative model with analytical geometric constraint enforcement via SDF projection significantly enhances the physical plausibility and quality of generated pose samples.
• A particle filter augmented with a belief-informed injection of tactile-conditioned hypotheses enables robust, online pose estimation, maintaining multimodal beliefs and converging effectively even with weak initial priors.

Summary by alphaXiv.